As an ER physician, my friend has often told me stories of his frustration with his AI-dependent residents. As he tells them, “Yes, I understand why you might think that was the problem, but it just does not look like that.” It would be easy to chalk it up to years of clinical experience. But what, really, is missing? What is it that clinicians see that AI does not?
So, I read with interest the article, “Intelligence without intuition: a mixed-methods pilot study on reasoning models in musculoskeletal physiotherapy for low-back pain.” As a relatively “late to the party” adopter of reasoning models (or AI), I wanted to know what AI does well, but also what it does not do well. The authors conducted a comparative analysis of reasoning models for clinical reasoning using five vignettes, three state-of-the-art reasoning models (Gemini 2.5, o3, and DeepSeek-R1-0528), and five experts in musculoskeletal physical therapy. They found that these reasoning models demonstrated “sufficient test–retest reliability and are competent or proficient in terms of their conceptual reasoning, completeness, correctness, relevance, and usefulness” when compared to expert reasoning. So, this is what AI appears to do well.
However, they found weaknesses in the reasoning models:
- Patient-centeredness – consideration of contextual factors such as the individual’s goals, treatment barriers, and psychosocial influences. Compared to expert reasoning, models “did not integrate lived experience and values” and “approached the patient as a diagnostic target rather than a person with agency.”
- Empathy – being “tuned in” to the patient emotionally. Compared to expert reasoning, “emotional engagement [was] absent” and there was “no therapeutic relationship.”
- Logical coherence – difficulty connecting reasoning steps to arrive at a conclusion. Compared to expert reasoning, there was a “strong biomedical tone that neglected alternative explanations” and reasoning was “overly rigid, with insufficient justification for sequential diagnosis.”
- And especially intuition – difficulty with experience-based pattern recognition or “clinical gut feeling.” Compared to expert reasoning, there was “rigid, data-driven reasoning, lacking moment-to-moment adjustment” and a lack of “dynamic hypothesis testing.”
So, this appears to be what AI does not do well. The authors noted that better prompting -encouraging greater emphasis on contextual factors, values, and patient emotions - might improve ratings for patient-centeredness and empathy. However, they could not recommend the use of AI in routine musculoskeletal practice due to its potential for reinforcing biomedical biases, its limited attention to individual differences, and the risk of plausible but inappropriate recommendations, especially in situations where experiential insight and intuition are essential.
Coming back to my friend, perhaps it now makes a bit more sense why he said, “it just does not look like that.” Perhaps he asked the patient about their preferences and included them in the decision. Perhaps he sensed the patient was distracted by the noisy ER and needed a quieter space to think clearly. Perhaps he realized that tachycardia could be a symptom of a seemingly unrelated diagnosis and needed further investigation to understand how it linked to other symptoms. And perhaps he “trusted his gut,” knowing that what he was seeing simply did not fit the assumed diagnosis.
So, let’s use AI for what it does well—and celebrate and develop what only clinicians can do.
Knauer R, Kalmring M and Rodner E (2026) Intelligence without intuition: a mixed-methods pilot study on reasoning models in musculoskeletal physiotherapy for low-back pain.Front. Digit. Health 8:1789412.doi: 10.3389/fdgth.2026.1789412
https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2026.1789412/full